Optimized short-term load forecasting in residential buildings based on deep learning methods for different time horizons

Abstract

The aim of this paper is to develop machine learning based framework to short-term load forecasting with high accuracy for residential building. The purpose is to develop a predictive model that assists energy companies in achieving a balance between energy consumption and generation by effectively managing the energy demand of consumers. The originality of this paper lies in two main parts. First, it analyzes effective relevant features such as time, calendar, and weather using a correlation matrix. Next, an optimized eXtreme gradient boosting model is employed to select key features, reducing the complexity of the training model. The second part proposes an intelligent parallel structure that utilizes gated recurrent units and convolutional neural networks for short-term load forecasting in different resolutions with minimal errors. The precise selection of hyperparameters significantly influences error prediction and accuracy. The metaheuristic-search-based algorithm, particle swarm optimization, is applied to find the optimal configuration for tunable parallel proposed model hyperparameters. The main results of this study demonstrate that the intelligent predictive model performs better than the latest models discussed in recent literature. The evaluation was conducted on a real-time time series dataset called Mashhad data. The model achieved impressive results in terms of standard metrics such as root mean square error of 44.28, mean absolute error of 29.32, mean absolute percentage error of 3.11 %, and R2 of 0.9229 % for the next 24 h. The benefits of this novel model play a crucial role in accurate short-term load forecasting, as well as in the decision-making and operations of power companies and smart grid.